[zeromq-dev] PUB/SUB on an epgm socket ... and CPU usage?

Steven McCoy steven.mccoy at miru.hk
Fri Jul 1 23:00:26 CEST 2011


On 29 June 2011 14:39, Ladan Gharai <lgharai at gmail.com> wrote:

> Sorry, I wasn't clear in my email - the high CPU usage is on the *send*
> side.
>
> But the good news is with nak_data_retries = 2,nak_ncf_retries=2 and
> ZMQ_RATE=200Mbps and an actual send-rate around 100Mbps  - I've had a pretty
> steady flow running for over a day now (sending and receiving at 100Mbps).
>>
>>
> For these data rates the CPU usage for the sender is at 56% and the
> receiver is at 30% - are these numbers normal for ePGM?
>
>
The rate engine can cause a lot of churn as OS sleep calls are not that
efficient yet.  This means some of that CPU usage can be taken away and you
will see no performance difference.

Hopefully high resolution timers make it into sleep calls sometime in the
future and such problems will be resolved.  The effect is more notable on
Windows which is why the code rounds up timers making the rate engine less
accurate.

If you run through oprofile you should get a good indication of where time
is being spent.

-- 
Steve-o
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.zeromq.org/pipermail/zeromq-dev/attachments/20110701/e36c6225/attachment.htm>


More information about the zeromq-dev mailing list